Re: [PATCH] net_sched: sfq: add optional RED on top of SFQ

[Rick Jones <rick.jones2@hp.com>]

On 01/06/2012 09:07 AM, Eric Dumazet wrote:
> Le vendredi 06 janvier 2012 à 17:56 +0100, Dave Taht a écrit :
>> On Fri, Jan 6, 2012 at 5:31 PM, Eric Dumazet<eric.dumazet@gmail.com>  wrote:
>>> Adds an optional Random Early Detection on each SFQ flow queue.
>>
>> netperf -t TCP_RR is useful
>> -t TCP_MAERTS will be interesting.
>> simultaneous ping?
>>
>
> I dont know what you expect from pings, since they are already coming in
> new flows (unless a ping flood is in effect), so RED doesnt fire for
> these packets.
>
> Same for TCP_RR : Since at most one packet is in flight per flow, RED
> cannot fire.

netperf nitpick :)  While I doubt that Dave Taht is running it that way, 
one can have multiple requests in flight on a single _RR test via the 
test-specific -b <additionaltrans> option.  That option is enabled by 
default (--enable-burst on the configure) in 2.5.0 and later.

netperf -t TCP_RR ...  -- -b 1

will cause netperf to have two transactions in flight at one time, -b 3 
will have four etc etc (actually it "slow-starts" to get to that level). 
  This can also be (ab)used to implement a single-connection, 
bi-directional throughput test such as I have in my "runemomni" scripts 
under doc/examples.  For example:

netperf -t TCP_RR ... -- -s 1M -S 1M -r 64K -b 12

As there is no select() or poll() call in the path, it is best to ensure 
that the SO_SNDBUF size on either end is large enough to hold 
[request|response]_size*simultaneous_trans at one time.

Depending on the size of the requests/responses one may want to add the 
test-specific -D option to set TCP_NODELAY.

happy benchmarking,

rick jones